Neural network algorithms for motion stereo

Abstract

Motion stereo infers depth information from a sequence of image frames. Both batch and recursive neural network algorithms for motion stereo are presented. A discrete neural network is used for representing the disparity field. The batch algorithm first integrates information from all images by embedding them into the bias inputs of the network. Matching is then carried out by neuron evaluation. This algorithm implements the matching procedure only once, unlike conventional batch methods requiring matching many times. The method uses a recursive least square algorithm to update the bias inputs of the network. The disparity values are uniquely determined by the neuron states after matching. Since the neural network can be run in parallel and the bias input updating scheme can be executed on line, a real-time vision system employing such an algorithm is very attractive. A detection algorithm for locating occluding pixels is also included. Experimental results using natural image sequences are given. >